Based on local tertiary mode without with reference to three-dimensional image objective quality evaluation method
Technical field
The present invention relates to a kind of stereo image quality evaluation methodology, especially relate to a kind of nothing based on local tertiary mode
With reference to three-dimensional image objective quality evaluation method.
Background technology
Since entering 21st century, along with reaching its maturity of stereoscopic image/video system treatment technology, and computer
The fast development of Networks and Communications technology, has caused people's tight demand to stereoscopic image/video system.Compare tradition haplopia
Dot image/video system, stereoscopic image/video system, owing to depth information can be provided to strengthen the sense of reality of vision, is given and is used
Family is more and more welcomed by the people with brand-new visual experience on the spot in person, has been considered as main the sending out of Next-Generation Media
Exhibition direction, has caused the extensive concern of academia, industrial circle.But, people are in order to obtain the most three-dimensional telepresenc and vision
Experience, stereoscopic vision subjective perceptual quality is had higher requirement.Stereoscopic vision subjective perceptual quality is to weigh axonometric chart
The important indicator that picture/video system performance is good and bad.In stereoscopic image/video system, gather, encode, transmit, decode and
The processing links such as display all can introduce certain distortion, and stereoscopic vision subjective perceptual quality will be produced in various degree by these distortions
Impact, the most effectively carrying out reference-free quality evaluation is the difficulties needing solution badly.To sum up, axonometric chart picture element is evaluated
Measure, and the foundation objective evaluation model consistent with subjective quality assessment is particularly important.At present, research worker proposes not
Few nothing for single viewpoint vision quality is with reference to evaluation methodology, yet with lacking Systems Theory further investigation stereoscopic vision perception
Characteristic, the most effectively without with reference to stereo image quality evaluation methodology.Compare single viewpoint vision quality without with reference to evaluating
Model, without needing to consider different type of distortion solid masking effect and associated with reference to stereo image quality evaluation model
Binocular competition/third dimension master factor impact on visual quality such as suppression and binocular fusion.It is thus impossible to simply existing
Single viewpoint vision quality is without being directly extended to without with reference in stereo image quality evaluation methodology with reference to evaluation model.Existing without ginseng
Examine assessment method for encoding quality and mainly carry out prediction and evaluation model by machine learning, but for stereo-picture, existing vertical
Body image evaluation method or the simple extension of plane picture evaluation methodology, do not consider binocular vision characteristic, therefore, how
In evaluation procedure, efficiently extract characteristic information, evaluation procedure carries out binocular vision characteristic combination so that objective evaluation
Result more conforms to human visual perception system, is to need that researchs and solves to ask during stereo-picture carries out evaluating objective quality
Topic.
Summary of the invention
The technical problem to be solved is to provide a kind of based on local tertiary mode without with reference to stereo-picture visitor
Appearance quality evaluation methodology, it can fully take into account stereoscopic vision characteristic such that it is able to be effectively improved objective evaluation result with
Dependency between subjective perception.
The present invention solves the technical scheme that above-mentioned technical problem used: a kind of based on local tertiary mode without reference
Three-dimensional image objective quality evaluation method, it is characterised in that its processing procedure is: first, the stereo-picture to distortion to be evaluated
Left view dot image and right visual point image implement Gauss gradient filtering respectively, obtain respective magnitude image and phase image, and
Calculate the anaglyph between left view dot image and the right visual point image of the stereo-picture of distortion to be evaluated;Secondly, according to treating
The left view dot image of stereo-picture of the distortion evaluated and the respective magnitude image of right visual point image and phase image, and left view point
Anaglyph between image and right visual point image, calculates the left and right viewpoint Feature Fusion figure of the stereo-picture of distortion to be evaluated
Picture;Then, use local tertiary mode operation that the left and right viewpoint Feature Fusion image of the stereo-picture of distortion to be evaluated is entered
Row processes, and obtains the upper mode image of its local tertiary mode and lower mode image;Afterwards, use statistics with histogram method respectively
Upper mode image and lower mode image are carried out statistical operation, and correspondence obtains the upper ideograph of the stereo-picture of distortion to be evaluated
As pattern image histogram statistical nature vector under histogram statistical features vector sum;Finally, standing according to distortion to be evaluated
Under the upper pattern image histogram statistical nature vector sum of body image pattern image histogram statistical nature vector, use support to
Amount regression forecasting obtains the evaluating objective quality predictive value of the stereo-picture of distortion to be evaluated.
This nothing comprises the following steps with reference to three-dimensional image objective quality evaluation method:
1. S is madedisRepresent the stereo-picture of distortion to be evaluated, by SdisLeft view dot image be designated as { Ldis(x, y) }, will
SdisRight visual point image be designated as { Rdis(x, y) }, wherein, 1≤x≤W, 1≤y≤H, W represent SdisWidth, H represents Sdis's
Highly, Ldis(x y) represents { Ldis(x, y) } in coordinate position be (x, the pixel value of pixel y), Rdis(x y) represents { Rdis
(x, y) } in coordinate position be (x, the pixel value of pixel y);
2. to { Ldis(x, y) } implement Gauss gradient filtering, obtain { Ldis(x, y) } magnitude image and phase image, right
{ G should be designated asL_dis(x, y) } and { PL_dis(x,y)};Equally, to { Rdis(x, y) } implement Gauss gradient filtering, obtain { Rdis(x,
Y) magnitude image } and phase image, correspondence is designated as { GR_dis(x, y) } and { PR_dis(x,y)};Wherein, GL_dis(x y) represents
{GL_dis(x, y) } in coordinate position be (x, the pixel value of pixel y), PL_dis(x y) represents { PL_dis(x, y) } in coordinate
Position is (x, the pixel value of pixel y), GR_dis(x y) represents { GR_dis(x, y) } in coordinate position be (x, pixel y)
The pixel value of point, PR_dis(x y) represents { PR_dis(x, y) } in coordinate position be (x, the pixel value of pixel y);
3. block matching method is used to calculate { Ldis(x, y) } and { Rdis(x, y) } between anaglyph, be designated as { ddis(x,
Y) }, wherein, ddis(x y) represents { ddis(x, y) } in coordinate position be (x, the pixel value of pixel y);
4. according to { GL_dis(x, y) } and { PL_dis(x,y)}、{GR_dis(x, y) } and { PR_dis(x,y)}、{ddis(x, y) }, calculate Sdis
Left and right viewpoint Feature Fusion image, be designated as { Fdis(x, y) }, by { Fdis(x, y) } in coordinate position be (x, the pixel of pixel y)
Value is designated as Fdis(x, y),,
Wherein, GR_dis(x+ddis(x y), y) represents { GR_dis(x, y) } in coordinate position be (x+ddis(x, y), the picture of pixel y)
Element value,PR_dis(x+ddis(x y), y) represents { PR_dis(x,y)}
Middle coordinate position is (x+ddis(x, y), the pixel value of pixel y), cos () is for taking cosine function;
5. local tertiary mode is used to operate { Fdis(x, y) } process, obtain { Fdis(x, y) } local ternary mould
The upper mode image of formula and lower mode image, correspondence is designated as { LTPU(x, y) } and { LTPD(x, y) }, wherein, LTPU(x y) represents
{LTPU(x, y) } in coordinate position be (x, the pixel value of pixel y), LTPD(x y) represents { LTPD(x, y) } in coordinate bit
It is set to (x, the pixel value of pixel y);
6. use statistics with histogram method to { LTPU(x, y) } carry out statistical operation, obtain SdisUpper mode image Nogata
Figure statistical nature vector, is designated as { HU(m)};Equally, use statistics with histogram method to { LTPD(x, y) } carry out statistical operation,
To SdisLower pattern image histogram statistical nature vector, be designated as { HD(m)};Wherein, { HU(m) } dimension be 1 × m' dimension, HU
M () represents { HU(m) } in m-th element, { HD(m) } dimension be 1 × m' dimension, HDM () represents { HD(m) } in m-th unit
Element, 1≤m≤m', m '=P+2, P represent the field parameter in the tertiary mode operation of local;
7. n is used " an original undistorted stereo-picture, set up it under different type of distortion difference distortion levels
Distortion stereo-picture set, this distortion stereo-picture set includes several distortion stereo-pictures;Then subjective quality assessment is utilized
Method evaluates the subjective scoring of the every width distortion stereo-picture in this distortion stereo-picture set, by this distortion axonometric chart image set
The subjective scoring of the jth width distortion stereo-picture in conjunction is designated as DMOSj;According still further to step 1. to step operation 6., with identical
Mode obtain the upper pattern image histogram statistical nature of every width distortion stereo-picture in this distortion stereo-picture set to
Amount and lower pattern image histogram statistical nature vector, by the jth width distortion stereo-picture in this distortion stereo-picture set
Under upper pattern image histogram statistical nature vector sum, pattern image histogram statistical nature vector correspondence is designated as { HU,j(m) } and
{HD,j(m)};Wherein, n " > 1, the initial value of j is 1, and 1≤j≤N', N' represent the distortion comprised in this distortion stereo-picture set
Total width number of stereo-picture, 0≤DMOSj≤ 100, { HU,j(m) } and { HD,j(m) } dimension be 1 × m' dimension, HU,jM () represents
{HU,j(m) } in m-th element, HD,jM () represents { HD,j(m) } in m-th element, 1≤m≤m', m '=P+2, P represent
The locally field parameter in tertiary mode operation;
8. using this distorted image set as training set;Then utilize support vector regression to all distortions in training set
Under the subjective scoring of stereo-picture and upper pattern image histogram statistical nature vector sum pattern image histogram statistical nature to
Amount is trained so that the error between regression function value and the subjective scoring that training obtains is minimum, and matching obtains optimum
Weighted vector WoptWith optimum bias term bopt;Followed by WoptAnd boptStructure obtains support vector regression training pattern;
Further according to support vector regression training pattern, to SdisUpper pattern image histogram statistical nature vector { HU(m) } and lower pattern
Image histogram statistical nature vector { HD(m) } test, it was predicted that obtain SdisEvaluating objective quality predictive value, be designated as Q, Q
=f (x),Wherein, Q is the function of x, and f () is function representation form, and x is input, and x represents
SdisUpper pattern image histogram statistical nature vector { HU(m) } and lower pattern image histogram statistical nature vector { HD(m) },
(Wopt)TFor WoptTransposed vector,Linear function for x.
Described step 2. in Gauss gradient filtering in scale parameter σ value be 0.5.
Described step 5. in the operation of local tertiary mode in field parameter P value be 8, local radius parameter R takes
Value is 1, in adaptive thresholding value matrix { T (x, y) } under be designated as that (x, (x, y) value is α × F to value T of element y)dis(x, y), its
In, α is intensity factor, takes α=0.05.
Compared with prior art, it is an advantage of the current invention that: by deep excavation stereoscopic vision perception characteristic, to be evaluated
The left and right viewpoint Feature Fusion image of stereo-picture of distortion carry out local tertiary mode operation, obtain its local tertiary mode
Upper mode image and lower mode image, then use statistics with histogram method respectively upper mode image and lower mode image to be carried out
Statistical operation, correspondence obtains pattern under the upper pattern image histogram statistical nature vector sum of the stereo-picture of distortion to be evaluated
Image histogram statistical nature vector, characteristic vector pickup method is simple, and computation complexity is low, due to the mistake to be evaluated obtained
The eigenvector information of genuine stereo-picture can preferably reflect the mass change situation of the stereo-picture of distortion to be evaluated,
That is: to the evaluating objective quality predictive value of stereo-picture of distortion to be evaluated can reflect that human eye regards exactly
Feel subjective perceptual quality, it is possible to be effectively improved the dependency of objective evaluation result and subjective perception.
Accompanying drawing explanation
Fig. 1 be the inventive method totally realize block diagram.
Detailed description of the invention
Below in conjunction with accompanying drawing embodiment, the present invention is described in further detail.
It is a kind of based on local tertiary mode without with reference to three-dimensional image objective quality evaluation method that the present invention proposes, and it is total
Body realizes block diagram as it is shown in figure 1, its processing procedure is: first, to the left view dot image of the stereo-picture of distortion to be evaluated and
Right visual point image implements Gauss gradient filtering respectively, obtains respective magnitude image and phase image, and calculates mistake to be evaluated
Anaglyph between left view dot image and the right visual point image of genuine stereo-picture;Secondly, standing according to distortion to be evaluated
The left view dot image of body image and the respective magnitude image of right visual point image and phase image, and left view dot image and right viewpoint figure
Anaglyph between Xiang, calculates the left and right viewpoint Feature Fusion image of the stereo-picture of distortion to be evaluated;Then, employing office
The left and right viewpoint Feature Fusion image of the stereo-picture of distortion to be evaluated is processed by portion's tertiary mode operation, obtains its office
The upper mode image of portion's tertiary mode and lower mode image;Afterwards, use statistics with histogram method respectively to upper mode image and
Lower mode image carries out statistical operation, and correspondence obtains the upper pattern image histogram statistics spy of the stereo-picture of distortion to be evaluated
Levy pattern image histogram statistical nature vector under vector sum;Finally, according to the upper pattern of the stereo-picture of distortion to be evaluated
Pattern image histogram statistical nature vector under image histogram statistical nature vector sum, uses support vector regression prediction to obtain
The evaluating objective quality predictive value of the stereo-picture of distortion to be evaluated.
The nothing of the present invention is with reference to three-dimensional image objective quality evaluation method, and it comprises the following steps:
1. S is madedisRepresent the stereo-picture of distortion to be evaluated, by SdisLeft view dot image be designated as { Ldis(x, y) }, will
SdisRight visual point image be designated as { Rdis(x, y) }, wherein, 1≤x≤W, 1≤y≤H, W represent SdisWidth, H represents Sdis's
Highly, Ldis(x y) represents { Ldis(x, y) } in coordinate position be (x, the pixel value of pixel y), Rdis(x y) represents { Rdis
(x, y) } in coordinate position be (x, the pixel value of pixel y).
2. use prior art to { Ldis(x, y) } implement Gauss gradient filtering, obtain { Ldis(x, y) } magnitude image and
Phase image, correspondence is designated as { GL_dis(x, y) } and { PL_dis(x,y)};Equally, to { Rdis(x, y) } implement Gauss gradient filtering,
Obtain { Rdis(x, y) } magnitude image and phase image, correspondence is designated as { GR_dis(x, y) } and { PR_dis(x,y)};Wherein,
GL_dis(x y) represents { GL_dis(x, y) } in coordinate position be (x, the pixel value of pixel y), PL_dis(x y) represents { PL_dis
(x, y) } in coordinate position be (x, the pixel value of pixel y), GR_dis(x y) represents { GR_dis(x, y) } in coordinate position be
(x, the pixel value of pixel y), PR_dis(x y) represents { PR_dis(x, y) } in coordinate position be (x, the picture of pixel y)
Element value.
In the present embodiment, the scale parameter σ in Gauss gradient filtering can value be σ=0.5.
3. existing block matching method is used to calculate { Ldis(x, y) } and { Rdis(x, y) } between anaglyph, be designated as
{ddis(x, y) }, wherein, ddis(x y) represents { ddis(x, y) } in coordinate position be (x, the pixel value of pixel y).
4. according to { GL_dis(x, y) } and { PL_dis(x,y)}、{GR_dis(x, y) } and { PR_dis(x,y)}、{ddis(x, y) }, calculate Sdis
Left and right viewpoint Feature Fusion image, be designated as { Fdis(x, y) }, by { Fdis(x, y) } in coordinate position be (x, the pixel of pixel y)
Value is designated as Fdis(x, y),,
Wherein, GR_dis(x+ddis(x y), y) represents { GR_dis(x, y) } in coordinate position be (x+ddis(x, y), the picture of pixel y)
Element value,PR_dis(x+ddis(x y), y) represents { PR_dis(x,
Y) in }, coordinate position is (x+ddis(x, y), the pixel value of pixel y), cos () is for taking cosine function.
5. existing local tertiary mode is used to operate { Fdis(x, y) } process, obtain { Fdis(x, y) } local
The upper mode image of tertiary mode and lower mode image, correspondence is designated as { LTPU(x, y) } and { LTPD(x, y) }, wherein, LTPU(x,
Y) { LTP is representedU(x, y) } in coordinate position be (x, the pixel value of pixel y), LTPD(x y) represents { LTPD(x, y) } in
Coordinate position is (x, the pixel value of pixel y).
In the present embodiment, the field parameter P value during locally tertiary mode operates is 8, local radius parameter R value is
1, it is designated as under in adaptive thresholding value matrix { T (x, y) } that (x, (x, y) value is α × F to value T of element y)dis(x, y), wherein, α
For intensity factor, take α=0.05.
6. use existing statistics with histogram method to { LTPU(x, y) } carry out statistical operation, obtain SdisUpper ideograph
As histogram statistical features vector, it is designated as { HU(m)};Equally, use statistics with histogram method to { LTPD(x, y) } add up
Operation, obtains SdisLower pattern image histogram statistical nature vector, be designated as { HD(m)};Wherein, { HU(m) } dimension be 1 ×
M' ties up, HUM () represents { HU(m) } in m-th element, { HD(m) } dimension be 1 × m' dimension, HDM () represents { HD(m) } in
M-th element, 1≤m≤m', m '=P+2, P represent the field parameter in the tertiary mode operation of local.
7. n is used " an original undistorted stereo-picture, set up it under different type of distortion difference distortion levels
Distortion stereo-picture set, this distortion stereo-picture set includes several distortion stereo-pictures;Then existing subjective matter is utilized
Amount evaluation methodology evaluates the subjective scoring of the every width distortion stereo-picture in this distortion stereo-picture set, and this distortion is three-dimensional
The subjective scoring of the jth width distortion stereo-picture in image collection is designated as DMOSj;According still further to step 1. to step operation 6.,
The upper pattern image histogram obtaining the every width distortion stereo-picture in this distortion stereo-picture set in an identical manner is added up
Characteristic vector and lower pattern image histogram statistical nature vector, the jth width distortion in this distortion stereo-picture set is three-dimensional
Under the upper pattern image histogram statistical nature vector sum of image, pattern image histogram statistical nature vector correspondence is designated as { HU,j
(m) } and { HD,j(m)};Wherein, n " > 1, as taken n "=3, the initial value of j is 1, and 1≤j≤N', N' represent this distortion stereo-picture
Total width number of the distortion stereo-picture comprised in set, 0≤DMOSj≤ 100, { HU,j(m) } and { HD,j(m) } dimension be 1
× m' ties up, HU,jM () represents { HU,j(m) } in m-th element, HD,jM () represents { HD,j(m) } in m-th element, 1≤m≤
M', m '=P+2, P represent the field parameter in the tertiary mode operation of local.
8. support vector regression (Support Vector Regression, SVR) is based on empirical risk minimization
New machine learning method and statistical theory, it can suppress over-fitting problem effectively, and therefore the present invention is by this distortion
Image collection is as training set;Then the support vector regression subjective scoring to all distortion stereo-pictures in training set is utilized
And pattern image histogram statistical nature vector is trained under upper pattern image histogram statistical nature vector sum so that pass through
Error between regression function value and subjective scoring that training obtains is minimum, and matching obtains the weighted vector W of optimumoptAnd optimum
Bias term bopt;Followed by WoptAnd boptStructure obtains support vector regression training pattern;Instruct further according to support vector regression
Practice model, to SdisUpper pattern image histogram statistical nature vector { HU(m) } and lower pattern image histogram statistical nature to
Amount { HD(m) } test, it was predicted that obtain SdisEvaluating objective quality predictive value, be designated as Q, Q=f (x),Wherein, Q is the function of x, and f () is function representation form, and x is input, and x represents SdisUpper
Mode image histogram statistical features vector { HU(m) } and lower pattern image histogram statistical nature vector { HD(m) }, (Wopt)T
For WoptTransposed vector,Linear function for x.
In order to verify feasibility and the effectiveness of the inventive method further, test.
Here, the stereo-picture of the distortion that analysis and utilization the inventive method obtains is carried out in employing LIVE stereo-picture distortion storehouse
Dependency between evaluating objective quality predictive value and mean subjective scoring difference.Here, assessment image quality evaluation side is utilized
The conventional objective parameter of 3 of method is as the Pearson correlation coefficient (Pearson under the conditions of evaluation index, i.e. nonlinear regression
Linear correlation coefficient, PLCC), Spearman correlation coefficient (Spearman rank order
Correlation coefficient, SROCC), mean square error (root mean squared error, RMSE), PLCC and
The accuracy of the evaluating objective quality predictive value of the stereo-picture of RMSE reflection distortion, SROCC reflects its monotonicity.
The objective quality utilizing the inventive method to calculate the every width distortion stereo-picture in LIVE stereo-picture distortion storehouse is commented
Valency predictive value, recycles existing subjective evaluation method and obtains every width distortion stereo-picture in LIVE stereo-picture distortion storehouse
Mean subjective scoring difference.Five will be done by the evaluating objective quality predictive value of the inventive method calculated distortion stereo-picture
Parameter Logistic function nonlinear fitting, PLCC and SROCC value is the highest, RMSE value the lowest explanation method for objectively evaluating objective
Dependency between evaluation result and mean subjective scoring difference is the best.The quality evaluation performance of reflection the inventive method
PLCC, SROCC and RMSE correlation coefficient is as listed in table 1.Knowable to the data listed by table 1, the distortion obtained by the inventive method
Dependency between the final evaluating objective quality predictive value of stereo-picture and mean subjective scoring difference is good, shows
Objective evaluation result is more consistent with the result of human eye subjective perception, it is sufficient to feasibility and the effectiveness of the inventive method are described.
Table 1 utilizes the evaluating objective quality predictive value of the stereo-picture of the distortion that the inventive method obtains to comment with mean subjective
Divide the dependency between difference